todo: every eval_gen, cache prefix on all processors and identify it by it's hashed string. all populations are then sent out with this hash

jarbus / evo

MIT License

1 stars 0 forks source link

todo: every eval_gen, cache prefix on all processors and identify it by it's hashed string. all populations are then sent out with this hash #117

Closed jarbus closed 1 year ago

jarbus commented 1 year ago

this no longer seems to be the bottleneck, running evals seems to be the bottleneck. 4-agents, 140 length episodes take 36 seconds per gen with batch size 2 and two groups